400 Section - Muscle activity based prediction of training progress in HAL-assisted walking for acute hemiplegic stroke patients

Objectives : We have previously reported improvements in gait following training with the lower-limb single-leg HAL (Hybrid Assistive Limb, Cyberdyne) in patients with acute-phase stroke and gait impairment (Watanabe et al., 2020; Puentes et al., 2018; Tan et al., 2020). In this study, we examined whether training progress during HAL-assisted gait rehabilitation can be predicted based on muscle activity recorded during training.

Design: The study included 12 patients in the acute phase of stroke who underwent HAL-assisted gait training. Each patient completed 9 training sessions, and surface EMG (Delsys Trigno Lab, 2 kHz) was recorded during gait with and without HAL in each session. Using an adaptation of our proposed EMG analysis method (Anastasiev et al., 2023), we extracted features from the vastus medialis and gluteus maximus muscles during the first training session under both with HAL and without HAL conditions. Classification accuracy for distinguishing high-response patients from others was evaluated using linear discriminant analysis with leave-one-subject-out cross-validation. High-response patients were defined as those whose Functional Ambulation Category (FAC) improved by ≥2 points.

Results: Among 7,381 extracted features, 300 achieved >90% accuracy (mean AUC-ROC 0.943 [range: 0.72–1.00]), and 99 achieved 100% accuracy. Four-dimensional features yielded higher accuracy than two-dimensional features.

Conclusion: The finding that features capturing differences between with HAL and without HAL conditions contributed to prediction accuracy is consistent with previous movement analyses, including muscle synergy studies. High classification performance obtained using low-dimensional features suggests the potential for a simple and practical prediction system. The ability to estimate the expected number of HAL training sessions may provide clinically useful information.